Self-Supervised Learning with Spatiality Preserving Representation for EEG Signals

ICLR 2026 Conference Submission17699 Authors

Keywords: EEG, brain-computer interface, self-supervised learning, representation learning

TL;DR: We introduce a novel SSL method with a coherence pseudolabel prediction task, teaching models to understand the intricate topographical organization of EEG signals and achieving SOTA results.

Abstract: Self-supervised learning (SSL) has revolutionized the field of deep learning with EEG signals, yet current approaches face a critical limitation: the loss of crucial spatial information due to architectures that fail to adequately preserve the one-to-one electrode relationship between the input and representation. To address this, we introduce Spatiality Preserving Representation (SPR) Learning. Unlike existing methods relying on reconstruction or temporal prediction with separate encoders, SPR learns spatial relationships through an innovative coherence pseudolabel prediction task, teaching models to understand the intricate topographical organization of brain signals that conventional approaches overlook. Through comprehensive evaluation, SPR demonstrates superior performance over state-of-the-art methods (4.7%, 9.7%, 1.6%, and 15.6% fine-tuning improvements over different datasets), learning meaningful spatial representations that capture the complex spatial-temporal dynamics inherent in EEG data. Our work opens new avenues for interpreting the relationships of different brain regions by prioritizing spatial awareness, and thus bridge the gap between functional connectivity analysis and self-supervised EEG representation learning.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 17699